Notifications represent one of the most powerful channels for user engagement, yet without smart targeting, they can quickly lead to high uninstall rates. The challenge isn’t whether to use notifications but how to target the right users at the right time with the right content. This article explores notification optimization strategies drawing from research published by Pinterest, Twitter, and LinkedIn on building intelligent notification systems at scale. The Double-Edged Sword The Double-Edged Sword Notifications serve a critical purpose. They onboard users to new features, retain existing users, and keep people informed about content that matters to them. They’re a direct line to users’ attention in an increasingly crowded digital landscape. But here’s the catch: that same direct access can backfire spectacularly. Users can disable notifications with a single tap, or worse, uninstall the app entirely. Once that trust is broken, it’s nearly impossible to recover. A broken link, an invalid promo code, or simply too many interruptions at the wrong time can turn one of the most powerful retention tools into the biggest churn driver. The Naive Approach: Optimize for Clicks The Naive Approach: Optimize for Clicks The intuitive approach is to build a predictive model to estimate P(click | notification sent), the probability that a user will click if you send them a notification. Use machine learning to predict which notifications will get the highest click-through rate, then send those. Note that the training data here only contains users who received the notification. This is the “Simple Click Model” approach, and it’s seductive in its simplicity. The metrics look great. Click-through rates are up and daily active users are rising. But there’s a fundamental problem with this approach: P(click | notification sent) The model optimizes for users who would have visited the app anyway. The most engaged users (the ones who check the app multiple times daily) will naturally have the highest predicted click rates. So the algorithm sends them even more notifications, potentially overwhelming them and degrading their experience. Meanwhile, the notification budget is wasted on people who didn’t need the nudge. You’re measuring correlation, not causation. High CTR doesn’t mean you created value - it might just mean you bothered the highly active users. The model optimizes for users who would have visited the app anyway. Enter Uplift Modeling: Thinking Counterfactually Enter Uplift Modeling: Thinking Counterfactually Here’s where things get interesting. Instead of asking “who will click?” we should ask “who will click because we sent them a notification?” because This is the core insight behind uplift modeling, sometimes called incremental value modeling. The mathematical formulation: Uplift = P(click | notification sent) - P(click | no notification) Uplift = P(click | notification sent) - P(click | no notification) We’re not just predicting behavior; we’re predicting the incremental impact of our action. This tells us which users are truly influenced by notifications, allowing us to: incremental impact Avoid wasting notifications on users who would visit organically Stop bothering users who won’t engage regardless Focus on the persuadable middle: users who need that nudge Avoid wasting notifications on users who would visit organically Avoid wasting notifications Stop bothering users who won’t engage regardless Stop bothering users Focus on the persuadable middle: users who need that nudge Focus on the persuadable middle Think of it like a marketing budget problem. A savvy company doesn’t spend advertising dollars on customers who will buy anyway, nor on those who’ll never convert. They target the persuadable: people whose behavior actually changes in response to the ad. The same logic applies to notifications. Visualizing the Four User Segments Visualizing the Four User Segments Users can be segmented based on their behavior patterns with and without notifications: Figure 1: User Segmentation for Targeting (Credit: Claude AI) Claude AI The uplift model identifies that top-left quadrant (users who need the nudge), allowing you to maximize incremental value while minimizing volume and annoyance. The Implementation Challenge The Implementation Challenge This sounds great in theory, but there’s a catch: any given user either receives a notification or doesn’t. You can’t observe both outcomes simultaneously. This is the fundamental problem of causal inference: we need counterfactual data. The solution? Run a randomized controlled experiment. Split users into two groups: Treatment group: Receives notifications Control group: Receives no notifications Treatment group: Receives notifications Treatment group Control group: Receives no notifications Control group This gives you training data for both scenarios, allowing you to estimate the true causal effect. Approaches to Modeling Approaches to Modeling The Two-Model Approach: Train separate models for the treatment and control groups where each model sees only the portion of the training data corresponding to its group (treatment vs. control). At inference time, evaluate each notification candidate using both models and then compute the difference. The larger the predicted uplift, the more valuable that notification. The Two-Model Approach: The Single-Model Approach: Combine all data and train one model where “notification sent” is a feature. During the inference stage, evaluate each example twice, once with treatment, once without by toggling the feature value and measure the difference. The Single-Model Approach: Both methods have their merits, though the single-model approach can generalize better since it learns from the full dataset. Figure 2: Uplift Modeling Approaches (Credit: Claude AI) Claude AI The Evaluation Problem The Evaluation Problem Here’s where it gets philosophically tricky: we can never observe the true uplift for any individual example. If we sent someone a notification and they clicked, we’ll never know if they would have clicked anyway. There’s no ground truth. This means traditional evaluation metrics don’t work. Instead, researchers use techniques like bucketing users with similar characteristics and comparing aggregate uplift values, or running ongoing A/B tests to validate model predictions at a cohort level. Beyond Short-Term Clicks: The Long Game Beyond Short-Term Clicks: The Long Game Even with uplift modeling, optimizing purely for immediate clicks misses a crucial dimension: long-term value. Pinterest’s research on this topic is particularly illuminating. Consider these scenarios: A user clicks the notification but bounces immediately. Did it create value or just annoy them? A user disables notifications after one too many interrupts. You’ve lost the channel forever. Notification arrives after other platforms have already shared the same content. It feels stale and redundant. A user clicks the notification but bounces immediately. Did it create value or just annoy them? A user disables notifications after one too many interrupts. You’ve lost the channel forever. Notification arrives after other platforms have already shared the same content. It feels stale and redundant. Training Models for Long-Term Value Training Models for Long-Term Value Pinterest’s approach focuses on delivering relevant, high-quality notifications by training models to maximize long-term value rather than immediate metrics. The key insight: analyze historical data to predict how sending a specific notification now affects user behavior and engagement over a longer time horizon. now This means asking not just “will they click?” but “will this notification contribute to sustained engagement over weeks and months?” A notification that drives a click today but leads to an unsubscribe tomorrow has negative long-term value, even if it looks successful in short-term metrics. To truly optimize for the long game, we need to model: Engagement Quality: Did the user spend meaningful time with the content, or just click and bounce? Incorporate engagement signals into the training labels to distinguish between superficial and meaningful interactions. Engagement Quality: Negative Signals: Build predictive models for notification unsubscribes and app uninstalls. Pinterest’s research examines user activeness three weeks after an unsubscribe event to build training data for these predictive terms. This delayed measurement captures the true long-term impact. If the user unsubscribed and disengaged, we’ve identified an undesirable notification pattern. Negative Signals: Content Freshness: In a world where users get the same news from multiple apps, being first matters. Being late is worse than not sending at all. LinkedIn’s research demonstrates this principle at scale. They built their notification system on stream computing infrastructure to deliver notifications within seconds of user activity. When a post starts trending or gains significant engagement, delivering that notification immediately provides recipients an opportunity to join the conversation while it’s still active. Their system uses constrained optimization to balance engagement and send volume while guaranteeing freshness, processing millions of real-time notification decisions daily. Content Freshness: Content Quality: Notifications from trusted, high-quality creators should be weighted differently than those from spammy or dubious sources. One clickbait notification can erode trust that took months to build. Content Quality: Effective notification models typically incorporate various types of features: Notification activity history: How has this user responded to past notifications? What types, topics, and timing patterns work for them? Notification activity history: App activity patterns: When do they typically use the app? How engaged are they overall? App activity patterns: User demographics and context: Location, time of day, and day of week matter enormously. A notification at 2am is worse than no notification. User demographics and context: Send volume limits: Beyond predicting which notifications to send, systems must also determine how many notifications each user should receive. Twitter’s research demonstrates a hybrid approach using grid search based on user cohorts to set personalized notification caps. Rather than applying a one-size-fits-all daily limit, they segment users into cohorts and optimize send volumes separately for each group, recognizing that power users might tolerate more notifications while casual users need stricter limits. Send volume limits: research Different user cohorts need different strategies, and the system needs to learn these patterns at scale. Moreover, the time horizon for notification decision making extends over several days as opposed to session-level optimization. This longer view is crucial for modeling cumulative effects like notification fatigue. The Irrecoverable Loss The Irrecoverable Loss Perhaps the most sobering aspect of notification strategy is that some mistakes can’t be undone. A user who uninstalls the app due to notification fatigue is unlikely to return. Unlike most product decisions that you can iterate on, aggressive notification strategies can cause irrecoverable loss. This asymmetry should inform strategy. When in doubt, send fewer high-quality notifications rather than more mediocre ones. Respect users’ attention as the finite resource it is. The Real Optimization The Real Optimization The notification paradox reveals a deeper insight about digital products: the tools that can drive the most engagement are also the ones that can erode it most quickly. Traditional metrics optimize for reach and clicks, but these measure activity, not value. Uplift modeling and long-term value frameworks offer a different lens, one that asks not just who will respond, but who truly benefits from the interruption. The irony is that sending fewer, better-targeted notifications often delivers more value than aggressive broadcast strategies. The irony is that sending fewer, better-targeted notifications often delivers more value than aggressive broadcast strategies. Platforms that treat user attention as a finite and precious resource tend to see better long-term retention. Those that over-rely on notifications risk higher uninstall rates and user disengagement over time. Thanks for reading! Feel free to reach out on LinkedIn with any feedback or questions. LinkedIn Disclaimer: The views expressed in this article are solely my own and are not affiliated with my employer in any way. In writing this, I have not utilized any proprietary or confidential information. Disclaimer References & Further Reading References & Further Reading Notification Volume Control and Optimization System from Pinterest: link Near Real-Time Optimization of Activity-based Notifications from Linkedin: link Optimizing push notifications decision making by modeling the future from Twitter: link Uplift modeling techniques: link1, link2 Notification Volume Control and Optimization System from Pinterest: link link Near Real-Time Optimization of Activity-based Notifications from Linkedin: link link Optimizing push notifications decision making by modeling the future from Twitter: link link Uplift modeling techniques: link1, link2 link1 link2
